系統管理
顯示結果 
搜尋替代 
您的意思是 

sar -d 裡面的% busy 是怎樣算出來的

sar -d 裡面的% busy 是怎樣算出來的

%busy的解釋是 Portion of time device was busy servicing a request;

但是這是怎樣算出來的呢?

EVA上的出現的100 % busy 與local disk 100 % busy

雖然都是100,但是基準點應該不一樣才是

sar 又是怎樣出分出EVA 與local disk 的i/o throughput是不同的,而來定義多少% busy呢?
3 回覆

sar -d 裡面的% busy 是怎樣算出來的

Hi,

How to define "Busy" ?

Which is this person is doing his job.



How to define a disk is busy ?

Which is this disk is servicing a request.



sar is a counter based tool, kernel keep

counting the device busy status periodicly.

(not sure what the sample period is, the

passable value is 100ms I guess).



For each sample time, kernel take a

a look on the device status, if this device is

in servicing a request, this counter will be

increased. sar is based on this counter to

get the value of %busy.



Consinder this senario, if system behavior as

bust I/O traffic in 10 sec and then no more

traffic on the follow on 90 sec, you

will get 100 %busy if the you do sar 1 10

But you will get 10 %busy if you do sar 1 100

Isn't it ?



By the way, there are patches for sar, sar do

inaccurate count without patch.



For 11.11 PHKL_33858 (Reboot : yes)

For 11.23 PHCO_32702 (Rebbot : no)



Hope this is help.



..

.



sar -d 裡面的% busy 是怎樣算出來的

您的回文提到

""""For each sample time, kernel take a look on the device status, if this device is in servicing a request, this counter will be increased."""

假設Local disk and EVA lun都接受相同的i/o動作,理論上我得到的servicing a request counter 值應該有相同,所以得到的busy的值應該也應該接近才是?

還是會因為EVA處理的比較快,counter當sar 去query 已經沒有servicing,所以值會比較低???(我會盡可能將sar 的lnterval調短來測試)

如過答案是前者的話

假設busy 都是100,我想EVA 應該還有餘力處理多餘的i/o才是,那是不是相同的busy 100對EVA來說,並不是真的已經到了max了

sar -d 裡面的% busy 是怎樣算出來的

I think if we put the same loading to local disk and EVA LUN, 因為EVA處理的比較快,counter當sar 去query 已經沒有servicing,所以值會比較低.



From the avserv filed of sar -d output, we can see the value of EVA disk is much better then local disk. In the meanwhile, the avque of of local disk is easy to queue up, but EVA is not.



One more step, donot forget the scsi queue depth, for local disk the scsi queue depth is limited on 8, but for EVA LUN, you can size it up for more, regarding EVA is more capable.

The experimental value is 32 depened on how many LUNs in the whole EVA. scsictl is tool to szie up the queue depth.



Just a query, have you check the patch level of sar ? The command is

# what /usr/bin/sar



..

.
//Add this to "OnDomLoad" event